Code World Model (CWM) 32B
32B dense code model mid-trained on Python-interpreter and Docker execution traces; 65.8% on SWE-bench Verified with test-time scaling.
Trains a coder on observation-action trajectories so it learns to simulate program execution ('world model for code'), then multi-task RL on verifiable coding, math and SWE environments; 131K context. Released checkpoints after mid-training, SFT and RL for research.
- Date
- Wednesday, 24 September 2025
- Lab
- Meta
- Kind
- open-weights
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified (pass@1, with test-time scaling) | 65.8% LiveCodeBench 68.6%, AIME 2024 76.0% | company |
License is research-oriented (exact terms not confirmed from the page opened). arXiv 2510.02387 appeared in October 2025.
Sources
- ai.meta.com/research/publications/cwm-an-open-weights-llm-for-research-on-code-generation-
- huggingface.co/facebook/cwm
This record was checked against its sources on 6 October 2026. How we check